An embedded multi-core stream architecture is developed to solve the intensive computation of real-time multi-target systems, such as GNSS/INS state estimation problems for multi-robot systems. A number of isomorphic stream cores are integrated in a single chip, interconnected with internal network. Each stream core employs 4 arithmetic clusters to support double precision floating point number operations. The data-path and control-path are both full pipelined, and result forward technology is used to solve the data conflict. Hierarchical storage structure, internal DMA controller and 64-bit internal buses can provide sufficient data bandwidth to feed the FALUs. Experimental result shows that more than 99% executing time is reduced for the EKF algorithm in GNSS/INS integrated system based on the proposed stream architecture, compared with other embedded processing platforms. An evaluation stream core is implemented in TSMC 65nm CMOS technology, the synthesis result shows the time delay is 4.94ns in the worst case, and the power consumption is 404.7 mW when it works at 200MHz.